All pull requests

Pull requests

• Fix default dense TP reduction in batch-invariant mode

  community-request
  #8321·Sunt-ing opened Oct 18, 2026Contributor·5 of 7 tasks·

  ·
  1

• Optimize MoE expert compute

  community-request
  #8320·Waterpine opened Oct 18, 2026·3 of 7 tasks·

  ·
  2

• Sum the GatedDeltaNet out_norm gradient across TP without sequence parallelism

  community-request
  #8319·12kimih opened Oct 18, 2026·1 of 7 tasks·

  ·
  1

  1

• Fix torch Inductor race condition in tests

  complexity: low
  #8318·tdene opened Oct 18, 2026Contributor·Queued·1 task done
  9

• fix(inference): promote TorchSampling temperature divisor to FP32

  community-request
  #8316·muditagrawal2007 opened Oct 17, 2026·

  ·
  1

  1

• fix(optimizer): skip redundant state reload in DistributedOptimizer.load_state_dict

  community-request
  #8315·muditagrawal2007 opened Oct 17, 2026·

  ·
  1

  1

• Add output-layer activation gradient clipping

  complexity: low
  #8314·mkhona-nvidia opened Oct 17, 2026Contributor·1 task·

  ·
  1

• MFSDP v2: add serial owner-compute Muon for 2D parameters

  complexity: medium
  #8313·wujingyue opened Oct 17, 2026Contributor·1 task done·

  ·
  7

• [MFSDP] Support shortcut MoE blocks and split residual units

  complexity: medium
  #8312·wujingyue opened Oct 17, 2026Contributor·3 of 6 tasks·

  ·
  1

• revert(multimodal): AUT-2553 remove obsolete VLM evaluation benchmarks

  Run full unit tests
  #8311·svcnemo-autobot opened Oct 17, 2026Collaborator·

  ·
  4

• fix(core): compare major version in experimental EOL check

  community-request
  #8309·NUAAwqy opened Oct 17, 2026·

  ·
  2

• Support GDN2 and TE CUDA graphs for recurrence offloading

  community-request
  #8308·Pinoeer-kingxi opened Oct 17, 2026·6 of 7 tasks·

  ·
  1

• Optimize NCCL MoE dispatch

  community-request
  #8307·Waterpine opened Oct 17, 2026·3 of 7 tasks·

  ·
  2

• Optimize NVLS MoE dispatch

  community-request
  #8306·Waterpine opened Oct 17, 2026·3 of 7 tasks·

  ·
  2

• chore: AUT-2547 nightly sync main into dev (2026-10-16)

  Run functional tests
  #8305·svcnemo-autobot opened Oct 17, 2026Collaborator·

  ·
  13

• build: use PyPI nvidia-cudnn-frontend 1.29.1 in CI dev image

  Approvedcomplexity: lowcore_r0.20.0
  #8303·ZhiyuLi-Nvidia opened Oct 17, 2026Contributor·

  ·
  1

• Refactor schedule node activation ownership

  #8302·vthumbe1503 opened Oct 16, 2026·3 of 7 tasks·

  ·
  1

• Add full_recompute_input to fine-grained activation offloading

  community-request
  #8300·avigyabb opened Oct 16, 2026·4 of 7 tasks·

  ·
  1

• Shard KPool indexer selection across tensor-parallel ranks

  community-request
  #8299·avigyabb opened Oct 16, 2026·4 of 7 tasks·

  ·
  1

• Add bf16 reduce-scatter option for 2-rank expert-TP MoE combine

  community-request
  #8298·avigyabb opened Oct 16, 2026·4 of 7 tasks·

  ·
  1

• Release router probs in the all-to-all dispatcher after combine

  26.12community-request
  #8297·avigyabb opened Oct 16, 2026·5 of 7 tasks·

  ·
  1

• Optimize sigmoid MoE inference routing with bit-exact outputs

  community-request
  #8296·Waterpine opened Oct 16, 2026·3 of 7 tasks·

  ·
  2

• Support standard DCP save/load with MFSDP v2 state-dict hooks

  #8295·wujingyue opened Oct 16, 2026Contributor·3 of 5 tasks·

  ·
  1

• Chenshuz/gigatoken integration

  complexity: medium
  #8294·jupiterepoch opened Oct 16, 2026Contributor·6 tasks done·

  ·
  3

• Fix duplicate parameter all-gathers across MoE distributed optimizers

  community-request
  #8292·ewan0x79 opened Oct 16, 2026·4 of 6 tasks·

  ·
  1